RESEARCH DOSSIER
Cyber Insurance Coverage of Artificial Intelligence Enabled Threats
Abstract. Artificial intelligence alters the capabilities of cyber adversaries, the exposure of organizations operating digital systems, and the evidence available for insurance underwriting and claims. This monograph examines the relationship between AI-enabled malicious activity and the contractual transfer of its financial consequences. It integrates selected primary threat intelligence, official crime reports, three authentic sample cyber insurance forms, judicial decisions, London market requirements, regulatory context, and established actuarial literature. It distinguishes malicious use of AI, attacks on AI systems, dependent-service disruption, and ordinary AI operational failure; compares discovery, interruption, and liability triggers; analyzes synthetic impersonation, extortion, attribution, and contract coordination; and develops reproducible hypothetical loss-adjustment, classification-error, and accumulation calculations. The synthesis establishes why the role of AI is an additional exposure and evidentiary dimension while indemnity continues to depend on the insured event and operative terms. The research uses a targeted documentary comparison, not a representative global empirical sample. Public evidence examined does not identify a calibrated AI-specific insurance loss distribution. The contribution consists of comparative analysis and reasoned synthesis; no new named theory or claim of priority is made.
Keywords: Cyber insurance; artificial intelligence; AI-enabled cyber threats; deepfakes; social engineering; computer fraud; cyber extortion; insurance contracts; cybersecurity; threat attribution; business interruption; underwriting; cyber accumulation; reinsurance; claims evidence; classification error; insurance risk.
Introduction
The economic object of this research is the transfer of financial loss associated with malicious digital acts whose execution, preparation, targeting, or persuasive power is materially assisted by artificial intelligence (AI). The institutional object is the cyber insurance contract, through which a party undertakes to finance specified consequences of such events. The two objects intersect when an attack reaches an insured exposure and produces a legally and contractually recognizable loss. AI changes the means by which a threat becomes effective; the policy determines the conversion of that effectiveness into an indemnity obligation.
The parent discipline, Cyber Insurance, establishes the wider scientific analysis of loss transfer, security economics, contract risk, and accumulation. This dossier concentrates on the additional difficulties created by AI-enabled threats. It takes the position that the most informative object of inquiry is the entire causal sequence from adversary capability through financial consequence and contract response. The analysis includes attacks that use AI against conventional systems and attacks directed at AI systems themselves, provided that the resulting event is relevant to cyber insurance. Legitimate system errors, professional negligence, product liability, and the operational risk of insurers are treated at the boundary where they interact with malicious cybersecurity exposures.
The study asks three questions. First, what can public evidence establish about AI-enabled attacks rather than merely forecast? Second, how do actual policy provisions, definitions, time triggers, and legal interpretations allocate their financial consequences? Third, how does uncertainty about AI involvement change underwriting, claims measurement, and the capacity to pool risk? The international comparison is deliberately asymmetric: the United States supplies public claims-market statistics and selected judicial opinions, the London market supplies institutional wording practice, Hong Kong supplies a documented incident, and the European Union supplies a regulatory context. These jurisdictions are analytically complementary and do not constitute a representative worldwide sample.
The evidence is a purposive documentary sample through November 18, 2026. It comprises accessible original threat and crime reporting, selected specimen policy wording, judicial texts in public reproductions, market documents, publicly available technical guidance, and actuarial literature. Several sources could only be checked at the abstract or indexed official-text level; the evidence ledger identifies that scope. No proprietary portfolio, individual claim file, confidential underwriting data, experimental attack measurements, or policy endorsements issued to a named victim were available. This design supports conditional coverage interpretation and qualitative comparison, together with reproducible hypothetical arithmetic. It does not support a calibrated AI-specific loss distribution or jurisdiction-wide legal conclusions.
Conceptual and Historical Foundations
The distinction between capability and insured peril
An AI-enabled threat is a malicious capability or activity in which statistical learning, generative systems, or related automated inference materially supports an adversary's actions. AI involvement can be direct, as in synthetic impersonation, or indirect, as in the analysis of information used to select targets. The classification should preserve the conventional peril: fraudulent instructions remain fraudulent instructions, unauthorized access remains unauthorized access, and coercive threats remain coercive threats. This dual description makes observations comparable over time even when the tools used to commit an offense change.
The history of cyber insurance is also a history of changing technological vocabulary. Independent risk transfer for information privacy, network security, interruption, and digital crime developed because ordinary property, crime, and liability contracts could leave substantial questions about intangible loss, computer-related causation, and the scope of covered systems. Modern AI is introduced into an existing contractual field. A generation of attack instruments can change while the contract's language remains constant. Alternatively, terms can be revised even when the mechanism of loss remains similar. A historical comparison must track both processes.
The development is observable in the transition from earlier computer-fraud litigation to contemporary specimen wording expressly contemplating synthetic impersonation. That transition has scientific significance: it shows the responsiveness of the definition of the insured event to technological substitution. It does not establish that every older contract excludes an event involving AI, because technology-neutral wording may already encompass the facts.
Several distinct roles of AI
Four situations merit separate description. In attacker-side use, the adversary employs AI while the insured organization operates conventional information systems. In attacks against AI, a model, its associated data, its interfaces, or the surrounding computing service is the affected asset. In dependency losses, an incident at a third-party provider of AI or supporting digital services interrupts the insured's operations. Finally, a legitimate AI system may generate harmful outputs or perform unauthorized actions in the absence of an external attacker. In the last situation, determining whether a cyber policy covers nonmalicious failures or whether another insurance line is relevant requires attention to the wording and liability basis.
These categories identify observable roles and can coexist. An organization that uses an AI-enabled workflow can receive a malicious input and suffer an unauthorized transaction; the adversary may also have used AI to prepare the input. A loss record limited to a binary AI flag collapses these distinctions. A scientific record should identify the affected system and the mechanism of the economic harm alongside the role attributed to AI.
Vassilev and colleagues (2025) organize adversarial machine learning terminology around machine-learning approaches, stages of the system life cycle, and adversary characteristics. The publication abstract and keywords identify classes including evasion, poisoning, privacy compromise, and abuse. These are technical categories of exposure. Their insurance significance turns on whether model integrity, confidentiality, availability, or an associated legal interest has been compromised, and whether that consequence is within an insured provision.
Evidence and intelligence assessments
The UK NCSC's (2025) assessment of threats through 2027 anticipates more effective intrusion operations and an expanded attack surface as organizations adopt AI. The assessment concerns anticipated cyber activity and explicitly focuses on intrusion. Its probabilistic language expresses intelligence judgment, rather than a measured insured-loss rate.
The economic inference is that an improvement in attacker productivity has several possible transmission mechanisms. It can increase the number of attempts per unit of time, improve the chance of success conditional on an attempt, reduce the cost of operating a campaign, or shift the mix of targeted organizations. Defensive improvements can alter the same quantities in the opposite direction. Consequently, the quantity an insurer wishes to estimate is the change in covered losses conditional on the joint evolution of attack, defense, exposure, and policy terms. Threat forecasts alone cannot identify this quantity.
A related historical problem is changing visibility. A sophisticated synthetic message may leave little reliable evidence of its origin. The growing ability to recognize AI content, greater media attention, and the introduction of reporting categories can all increase recorded AI involvement without an equivalent increase in underlying incidence. A longitudinal series therefore requires definitions and detection methods to be versioned. Apparent discontinuities near the introduction of a descriptor should be examined as possible measurement artifacts before they are attributed to a change in the hazard.
Primary Documents and Scope of AI-Assisted Threats
Contracts allocate consequences through operative language
A standalone cyber policy is a collection of conditional promises. Coverage must be examined through the declared insured, the insuring agreement actually purchased, the operative event definition, its temporal and territorial requirements, the defined loss, exclusions, conditions, and monetary restrictions. A policy heading is evidence of organization, rather than a sufficient statement of coverage. The same incident may generate response expenses, lost net income, theft, and third-party liability, each governed by a different trigger.
At-Bay's (2023) cyber policy form separates privacy and network response, interruption, extortion, financial fraud, and media coverage. Its example of interruption requires a system disruption beginning in the policy period, whereas several other first-party promises turn on discovery and liability promises on claims made. Its aggregate and sublimit provisions constrain the sum of recoverable amounts. This documentary observation establishes that assigning a single trigger to the whole cyber policy would produce misclassification.
The contrast becomes sharper in the treatment of deception. AXIS's (2020) specimen crime definitions address both security-incident-driven alteration of information and deceptive employee payment instructions. The latter branch contains an authentication-related restriction in the definition of covered crime loss. The technology behind an apparently valid request does not by itself establish whether that restriction was met.
The Coalition CIC sample policy expressly encompasses deepfake impersonation within its definition of funds transfer fraud. Its discovery-based agreement and its separate treatment of transfer liability connect different insured interests to different evidence. The sample also provides for a distinct $250,000 limit and $2,500 retention for the relevant fraud agreement. These are sample declarations, not estimates of typical market terms.
An explicit deepfake reference reduces uncertainty about one feature of the communication. It leaves open other eligibility questions, including the parties involved, the insured's interest in the transferred assets, the timing, loss measurement, and the continuing effect of exclusions. In a claim comparing an unmodified older form with a new form, the relevant counterfactual is the entire revised contract, not the mere presence of the word deepfake.
Comparative documentary matrix
| Document | Relevant trigger or definition | Material question in AI-enabled deception |
|---|---|---|
| At-Bay AB-CYB-001.2 (2023) | Fraudulent instructions accepted and discovered within the period; separate computer-crime grant | Which defined instruction or crime caused which loss? |
| AXIS 1012561 0120 specimen | Employee reliance or security-related fraud; restriction concerning authentication | Which mechanism applies and what independent verification occurred? |
| Coalition CIC sample (undated) | Deepfake impersonation expressly contemplated; discovery-based transfer grant | Is the impersonated person or organization within the definition, and whose loss was incurred? |
The samples were chosen for accessible operative wording and analytical contrast. They do not identify market shares, current availability in a particular state or country, endorsements attached to issued contracts, or standardized policyholder rights. Future comparative studies should sample filed or issued versions by jurisdiction, effective date, insurer, size segment, and endorsement. They should record the declaration and manuscript endorsement together with the base form.
Time, third parties, and AI-specific limitations
Suppose fraudulent communication is prepared in December, delivered in January, acted on in February, found in March, and reported in April. Each date may represent a different legally material fact. A discovery trigger may attach in March, a claims-made provision may attach when a third-party demand is made, and an interruption provision may depend on a disruption that began earlier. Related-incident provisions may relate certain losses back to an earlier event. A change of insurer or limit between the stages makes the chronology particularly important. The analyst should avoid deciding the coverage year merely by dating the AI-generated message.
AI services expand the practical importance of identifying systems within a contract. An enterprise may rent a hosted model, license software, maintain a private application, and outsource maintenance to several parties. Event coverage for its own network and contingent interruption coverage for another organization's system may involve different definitions and limitations. The location of data, contractual rights of access, and the location of the service provider's operations are separate facts. A broadly worded computer-system definition may encompass AI software, while a contingent-interruption provision might still be limited to qualifying providers or types of disruption.
The LMA's AI loss scenarios survey describes a model-wording review that treated AI as a category of software for the purpose of the reviewed cyber definitions. Its scenarios and responses were exploratory, with a variety of views and substantial acknowledged data limitations. That supports examining coverage through actual definitions rather than presuming a novel peril is excluded simply because it has a novel name.
Where an endorsement excludes or limits AI-specific exposures, its relationship with the underlying grant requires exact interpretation. An exclusion referring to the insured's development or supply of a model may have a different reach from one referring to any loss arising from use of AI. Exceptions, write-backs, and causation words are economically material. A narrow endorsement can delimit a recognizable activity; a broad causal exclusion can make the coverage outcome dependent on evidence that may never be obtainable from an adversary. Contractual precision has a direct effect on selection, loss-adjustment costs, and the degree of risk genuinely transferred.
Mechanisms and Evidence for the Insured Event
Impersonation and the reliability of authorization
In synthetic impersonation, an image, voice, or message transmits an apparent identity. Economic authority to transfer assets exists separately from that appearance. The distinction is especially important where the employee who controls the payment system actually authorizes a transaction after being misled. The technical network may have performed as designed, even though the organizational decision was induced by deception.
A documented incident illustrates the distinction. The Hong Kong government's (2024) Legislative Council deepfake reply recounts a police investigation involving a prerecorded conference, mistaken authorization of payments to five local accounts, and a loss of about HK$200 million. The reply does not establish an insurance claim or payment. Its evidentiary value is to demonstrate the combination of synthetic appearance, employee authorization, and large financial outflow in an actual reported incident.
The insurance research question arising from that event is the mapping of facts to possible provisions. The financial loss is the amount irrecoverably transferred, after appropriate accounting for bank recovery. A fraud provision might respond even without unauthorized access; a network-security grant might require additional facts. A third-party liability claim would involve a separate party with standing and an asserted legal obligation. Where a form requires verification, the record must establish what verification was prescribed and actually done. Calling a medium synthetic does not answer any of these questions.
Verification systems should establish authorization through a channel whose trust does not rest solely on content produced within the suspect interaction. For insurance purposes, the significant distinction is between a control represented in an application, a control technically deployed, a control used in the particular transaction, and a control that was demonstrably effective in preventing the loss. These states are not interchangeable. Documentation of an exception to a routine procedure is therefore as important as the documented procedure itself.
Intrusion, extortion, and interruption
In AI-assisted intrusion, financial consequences may result from compromise of data, loss of system availability, business interruption, recovery expenditure, or liability. AI's role may be supported by telemetry, adversary statements corroborated by independent evidence, recovered artifacts, or reasonable inference. Confidence should be recorded rather than treating the conclusion as certain. An ordinary intrusion can be assigned its security-event category even if its AI involvement remains unknown.
An extortion demand is also a distinct economic event. A threat can produce response expenditure without a ransom payment; nonpayment can coexist with substantial restoration costs. Payment can be a poor proxy for the severity of an insured's business disruption. The decision to pay or refuse involves legal restrictions, insured consent requirements, recovery options, human safety considerations where applicable, and the expected consequences. Coverage for extortion expenses does not transform an unlawful payment into a lawful transaction.
The U.S. Treasury's OFAC (2021) ransomware sanctions advisory identifies sanctions risks for participants facilitating payments, including cyber insurers and response firms, and specifies its own explanatory rather than legislative status. Its role in this study is to distinguish the potential contract grant from the permissibility of a proposed transaction. Determining permissibility for a particular claim requires the applicable binding rules and current facts.
Attacks on AI systems and operational loss
A model or integrated application can be affected by manipulation of inputs, unauthorized access to data, misuse of interfaces, or compromise of surrounding infrastructure. Financial loss can arise from investigation, containment, repair, interruption, or liability for information exposure. Some forms of model manipulation may produce a faulty output with no measurable deterioration of the underlying service; others can disable entire business processes. Consequently, measuring a technical anomaly by the number of abnormal outputs alone cannot establish the amount of interruption loss.
A provider may restore access while the customer still has to establish whether data or model integrity can be trusted. The date of technical restoration and the date of economically useful recovery may therefore differ. Indemnity depends on the definition of the restoration period and the documented necessity of the relevant steps. The analyst must distinguish new development or improved capacity from restoration to a defensible preincident state. Costs incurred to transform the insured's system architecture may be beneficial yet fall outside the coverage grant.
Investigating AI involvement
The appropriate evidentiary unit is an incident record with unique identifiers, provenance, a chronological event sequence, systems affected, persons authorized, communication channel, monetary transactions, response expenditure, and insurance provisions. Records should distinguish observed facts from investigator inference and from uncorroborated assertions. Classifications such as confirmed, supported but unconfirmed, suspected, and unknown are useful operational descriptions when the evidence criteria are stated. The absence of observable markers is not proof that an attacker did not employ AI.
Digital evidence can be fragile, especially in cloud services with short retention periods or after a system is restored. Original material, access logs, communication metadata, bank records, time zones, and investigator notes should be retained through documented custody procedures. The economic investigation can proceed when AI attribution is uncertain if the relevant event and loss are otherwise established. A contract that makes AI presence decisive will increase the importance, cost, and potential contestability of such attribution.
A claim examiner should test alternative explanations. For example, an apparent voice clone may have been a human impersonator, a recording, or ordinary audio manipulation. The payment may have followed a compromised vendor's email rather than the apparent executive's message. The ultimate loss may be shared with another organization or partly reversed. These possibilities matter even if the original attack narrative remains broadly credible. The proof required for insurance is the proof needed for the actual operative language and relevant law, not the strongest possible description of technological complexity.
Comparative Measurement of Coverage Interpretation
Judicial reasoning and factual variation in the United States
Earlier email-fraud decisions provide a documented legal comparator for technologically assisted deception. They are valuable because they separate the function of the technology from the language defining covered computer fraud. They do not decide contemporary AI-specific provisions.
In Medidata Solutions v. Federal Insurance (2018), the Second Circuit upheld the result for the insured in a summary order examining particular provisions about data entry or alteration. The court treated the manipulation of the receiving email system as sufficient under that wording despite the absence of conventional hacking. Its reasoning concerns the attack's actual effects and specific contract, and the decision's summary-order status limits its precedential significance.
The Sixth Circuit's published American Tooling Center decision (2018) reversed the judgment for the insurer and ruled for the insured at summary judgment in a fraudulent email payment case. The opinion addresses competing constructions and recognizes the significance of other, narrower contract wording. These decisions supply a basis for examining whether contemporary AI-mediated communications implicate the same provisions. A voice-only synthetic request, a fabricated video, a manipulated email header, and an unauthorized alteration of an account balance can be different events under a single computer-fraud definition.
A comparative legal study should specify the governing law, court hierarchy, procedural posture, policy form and language, relevant factual findings, disposition, and limits of the holding. A citation to the case caption alone is insufficient. The applicable law may use different approaches to proximate and direct causation, concurrent causes, ambiguity, and insurer defenses. The transfer of precedent across jurisdictions therefore calls for explicit reasoning and independent verification rather than translation of an outcome into a universal rule.
Complaint data and insurance data
The FBI's (2026) 2025 IC3 annual report records 22,364 complaints containing AI-related information and adjusted losses of $893,346,472. Its tables report 135 business email compromise (BEC) complaints in the AI-related subset, with $30,256,592 in associated losses. Investment complaints in that subset account for $632,041,188 of reported losses. These are complaint statistics; the descriptors reflect reported information and the process by which the agency classifies submissions.
One can calculate a descriptive loss-to-complaint quotient of approximately $39,945 by dividing the total AI-related adjusted loss by the corresponding reported complaint count. It is not modeled severity per insured incident, because the units include heterogeneous categories, the amount can be zero, the population lacks a policy-exposure denominator, and reporting may depend on severity and recognition. Nor is the BEC quotient an actuarial claim estimate. The highly different case mix is a reason to avoid interpreting an unsegmented quotient as a transferable severity parameter.
The NAIC (2026) cybersecurity insurance market report provides a different perspective. Its calendar-2025 analysis reports total loss ratios of 48.7% and 53.0% for 2024 and 2025 respectively. Its separate illustrations of account-size differences draw on other carrier and market sources. The change in the aggregate ratio is 4.3 percentage points. That is an observed market statistic within the report's accounting scope, not a causal estimate of AI effects.
Coalition's (2026) public cyber claims findings describe experience from more than 100,000 global policyholders during calendar 2025 and report a substantial rise in initial ransom demands. The figures refer to a provider-selected insurance population and an indicated observation window. The underlying individual policy and incident data were unavailable to this study. Coalition data reported again in the NAIC document should not be counted as independent evidence of the same effect.
Denominators, selection, and unknown attribution
A measurement comparison requires an observation unit and denominator that correspond. Threat reports may count campaigns, malicious communications, or investigated incidents. Law enforcement may count complaints. Insurers may count notices, claims, coverage agreements triggered, insured entities, policy years, or paid claims. Several claims can result from one attack, and several notices can concern one insured event. A firm may suffer multiple events while holding one policy. A portfolio may change its size, segmentation, terms, or selection practices during the observation period.
For a defined exposure of E policy-years and N recognized claims, crude frequency is N/E. That ratio estimates frequency only for the specified portfolio, period, eligibility and reporting convention. The numerator should specify zero-payment claims and open claims. Paid loss divided by the number of closed paid claims measures a particular conditional severity; incurred amounts with reserves estimate something different. Dividing criminal losses by insured premiums would combine incompatible populations and units.
AI attribution introduces an additional observation process. Let q be the true share of incidents with material AI involvement in a specified exposure, s the sensitivity of identifying that involvement, and f the false-positive rate. Under a simplified homogeneous classification assumption, the observed flagged share is sq + f(1 − q). If s and f are unknown, the true share cannot generally be recovered from the observed flag rate. If s exceeds f and both are established by validation, the expression can be rearranged; uncertainty in either classification parameter must propagate to the result. The relationship is ordinary measurement-error arithmetic and is not proposed as a new model.
Consider a hypothetical collection of 1,000 events with a 20% true share, a 60% sensitivity, and a 5% false-positive rate. The expected flagged count is 120 + 40 = 160, or 16%. If sensitivity rises to 80% while underlying prevalence and false-positive rate remain constant, the expected flagged count becomes 200, or 20%. The relative increase in recorded flags is 25% despite no actual change in prevalence. This example explains why a trend in labeled records alone cannot quantify the development of the underlying threat.
In practice, the assumption of homogeneous detection may fail. Large claims receive more forensic attention than small claims; certain fraud methods are more readily recognized than others; and the likelihood of reporting may depend on the policyholder's size and obligations. An analyst estimating differential AI involvement should stratify by event type and severity and validate the classification process on representative samples. Unknown cases must remain in the denominator where appropriate; deleting them can amplify selection bias.
Research design for causal questions
The effect of AI on insured outcomes can be described as a counterfactual comparison for the same exposure under different attack and defense conditions. This effect is difficult to identify because adoption, security maturity, sector, geography, attack opportunities, and policy terms change together. A simple before-and-after comparison would require a credible assumption that other determinants did not change or were properly controlled.
Potential study designs include matched cohort studies with consistent policy terms, time-to-event analysis with versioned exposure attributes, and event studies around independently dated attacker capability or defense deployments. Each design has limitations. Matching on observed characteristics leaves unobserved differences; a deployment date may coincide with other security upgrades; and controls may be adopted in anticipation of known threats. A robust study would state its identifying assumptions, include falsification tests, investigate missingness, and separate true security outcomes from insurance-selection effects.
Because this dossier lacks a linked exposure-incident-claim dataset, it makes no causal coefficient claim. The scientific result here is the specification of what such a coefficient would have to measure and why currently examined aggregate sources cannot identify it.
Financial Legal and Institutional Consequences
Loss adjustment and the effect of coverage constraints
Consider a policy covering one transfer fraud incident with an applicable limit of $250,000 and a deductible of $2,500. The insured's gross transferred amount is $400,000 and the bank recovers $100,000. Assume the remaining $300,000 is eligible loss, no other insurance applies, and the deductible reduces eligible loss before the limit. Then the indemnity is min($250,000, max($300,000 − $2,500, 0)) = $250,000. The insured retains $50,000 of the post-recovery loss. These figures are a synthetic calculation using illustrative contract terms; they are not an actual claim outcome or a statement of coverage under any complete policy.
If an investigation instead finds that only $200,000 of the net loss qualifies, indemnity becomes $197,500 and retained net loss becomes $102,500. If an exclusion entirely applies, the covered amount may be zero. The same technological event can therefore yield different settlements due to legal construction, recoveries, and limits rather than a different attack severity.
A general single-grant calculation, subject to its simplifying assumptions, is I = min(A, U, max(C − D, 0)), where C is eligible covered loss after defined recoveries, D is a per-event deductible, U is the remaining applicable sublimit, and A is the remaining overall aggregate. Actual forms can allocate deductibles, defense costs, sublimits and recoveries differently. A separate retention for a second grant cannot be inserted mechanically without determining the contract's treatment of related events. A full loss adjustment must also identify whether response expenses erode the same limit as theft or interruption.
Interruption loss requires a different counterfactual. Assume a firm has baseline daily contribution margin of $50,000, actual daily contribution margin of $15,000 after an incident, a six-day restoration period, a one-day time deductible, and $40,000 of independently eligible extra expense not already incorporated in margin. With uniform daily conditions and an assumed deductible that removes the first day's margin loss only, covered loss before monetary terms equals 5 × $35,000 + $40,000 = $215,000. Customers lost permanently may produce additional economic harm, but whether that loss falls within the contractual period and measurement definition would require separate analysis. A calculation of lost sales without deducting expenses saved could systematically overstate the indemnity.
Expected loss, dependence, and capital
Awiszus and colleagues (2023), in modeling and pricing cyber insurance, distinguish risks arising at individual insureds, common factors across insureds, and contagion or strategic interactions. They identify sparse data, evolving exposures and dependence as important limitations on classical methods. The distinction has a natural application to AI-enabled threats: increased activity across otherwise separate adversaries can produce a change in a common risk factor, while a compromise of a shared digital dependency can produce correlated or cascading losses.
Assume a hypothetical portfolio of 1,000 identical policies. Annual claim frequency per policy is 0.04 and expected capped loss per claim is $80,000. Under independence between the number and size of claims and a constant exposure, expected annual indemnities are $3.2 million. If frequency increases by 25% to 0.05 while mean indemnity remains fixed, expected indemnities rise to $4 million. If defensive measures instead reduce mean eligible capped loss to $60,000, expected annual indemnities become $3 million at the higher frequency. These calculations exclude expenses, profit, capital, taxes, reinsurance and investment income. They illustrate sensitivity, not an AI surcharge or recommended price.
Correlation can radically change a portfolio without changing its expectation. Suppose 1,000 policies each have a 1% probability of a $100,000 payment in a specified year. Under independent Bernoulli triggers, mean portfolio loss is $1 million and variance is 1,000 × 0.01 × 0.99 × $100,000² = $99 billion in squared dollars. Standard deviation is approximately $314,643. If all payments occur together with probability 1% and no payments occur otherwise, mean remains $1 million but the possible loss is $100 million; standard deviation is approximately $9.95 million. Perfect dependence is an extreme illustrative assumption. The comparison demonstrates why the expected premium cost does not describe risk capital needs.
In another synthetic scenario, a provider-dependent event affects 120 insured firms, each generating $800,000 of eligible indemnity. Gross portfolio loss is $96 million; an event excess-of-loss treaty with $50 million limit excess of $20 million would pay $50 million if the event meets its contractual aggregation and coverage terms, leaving $46 million net before any other recoveries or costs. If the treaty instead regards the circumstances as several unrelated events below attachment, the modeled recovery could be zero. The same portfolio losses can produce dramatically different net exposure because the treaty's event definition and underlying policy treatment determine attachment.
Underwriting evidence and security controls
Underwriting seeks to determine the insured's exposure, capability to prevent and contain losses, and dependencies. AI-related questions are most useful when tied to measurable organizational functions: the location of applications and data, permitted uses in sensitive workflows, authority to initiate financial transactions, logging and monitoring, third-party dependency, recovery testing, and the governance of exceptional permissions. A checkbox indicating use of AI has limited interpretive value without these details.
The functions of the NIST cybersecurity framework described by Pascoe and colleagues (2024) in CSF 2.0 support organizing cybersecurity governance and response. In underwriting analysis, a control description should distinguish a corporate policy from evidence of implementation, coverage across systems, and testing. A control may be associated with an improved underwriting outcome even if its causal effect on claim frequency is unproven. Eligibility criteria, contract conditions, pricing variables, and recommendations for improvement should remain clearly distinct.
Premium adequacy requires estimating both expected covered indemnity and the expenses of operating the insurance mechanism, including investigation and defense, risk capital compensation, acquisition costs and reinsurance where applicable. An insurer may revise terms or risk selection when evidence does not support a credible rate change. The uncertainty created by incomplete attribution may itself increase the variance of outcome and the cost of claims handling. A uniform surcharge based solely on AI use can obscure divergence between defensive use, attacker use, and high-dependency operational workflows.
Risk control also changes after coverage begins. A transaction approval method may become outdated; a provider can introduce new functionality; the insured can expand a model's permissions. Exposure assessment therefore benefits from dated snapshots and a clear account of material changes, subject to contract terms and local law. The insurer's access to evidence must be proportionate and respect confidentiality. Persistent access to unnecessary sensitive prompts, model inputs or customer data can generate a new concentration of privacy and operational risk.
State attribution, sanctions, and contract coordination
The London market's Market Bulletin Y5381 (2022) illustrates institutional action in response to systemic exposure from state-backed cyber activity. Its requirements apply to specified standalone cyber business at relevant inception or renewal dates, with a basis for attribution and provisions addressing significant impairment of a state. The LMA's cyber war clause classification distinguishes variants with different geographic and impairment treatment under the market's subsequent guidance. These documents establish market-specific requirements and classification, rather than a comprehensive statement of English or international insurance law.
AI-assisted activity does not imply state backing. The relevant evidence may include technical findings, official assessments, and the attribution mechanism selected by the contract. The analyst must distinguish observed technical artifacts from the identity of the entity acting behind them and from legal conclusions about state responsibility. A state-related incident may generate covered and excluded losses according to different clauses and locations. Broad conclusions about exclusion should not be drawn from the severity of the attack alone.
Insurance coordination arises because a single fact pattern can implicate crime coverage, standalone cyber insurance, property coverage, media liability, and technology errors-and-omissions. Coordination calls for identification of the claimant, the legally compensable interest and the amount under each grant. Other-insurance provisions, contribution, subrogation, and applicable law determine whether and how amounts are allocated. Double counting is a portfolio measurement error when the insurer group consolidates several covers on the same underlying loss without distinguishing separate indemnity obligations.
Regulatory and cross-border dimensions
The EU's Digital Operational Resilience Act (DORA) is an important context for financial entities. ESMA's DORA implementation overview identifies January 17, 2025 as its application date. This study uses the official overview only for that narrow institutional fact; the controlling consolidated legislative text could not be independently read back and detailed article-level legal conclusions are withheld. The general distinction remains analytically necessary: the obligations a financial institution has to manage its own digital resilience differ from its contractual obligations as an insurer of a third party's risk.
The insured may also face legal obligations to consumers, data subjects, counterparties, shareholders, or government agencies. The loss resulting from such obligations depends on jurisdiction, the type of obligation, facts of the incident, and policy definitions. A fine, contractual credit, voluntary payment or remediation cost should not automatically be treated as covered damages. Cross-border claims can involve several legal systems and several entities with different contractual interests. The controlling law for data protection, lawfulness of a payment, and interpretation of insurance can be different. A comparative analysis must record the source of each obligation rather than infer a global definition of liability from a model policy.
A full legal assessment for an individual incident would require the executed contracts, current legal authorities and relevant facts. The present comparison is an institutional and documentary scientific analysis. Its value lies in exposing the areas where cross-border variation changes the economics of risk transfer, rather than asserting identical contractual outcomes across insurance markets.
Integrated Scientific Synthesis and Research Agenda
What AI changes and what insurance must measure
The findings support a layered analysis of the effects of technological change. Attacker productivity can change the hazard. Deployment of AI systems can change the assets and dependencies exposed to that hazard. Detection and evidence quality determine observability. The insurance contract selects the consequences eligible for transfer. Portfolio structure and reinsurance determine the concentration of those consequences on an insurer's balance sheet. The five levels are distinct descriptive elements of an integrated analysis, and no new formal theoretical framework is claimed.
A core inference is that two otherwise identical insured losses can present different evidence of attacker tooling with no difference in contractual outcome when the relevant grant is technology-neutral. In contrast, two factually similar acts of impersonation can lead to different outcomes when one contract explicitly includes synthetic deception and another limits employee-induced transfer losses to a verified procedure. These comparisons explain why insurance research must examine the wording and economic consequence of the incident in tandem with security intelligence.
The role of AI can be incidental to a grant and materially relevant to risk estimation at the same time. An insurer may expect higher loss frequency because AI-assisted deception increases attack volume, yet pay a particular covered claim without determining whether AI was used. Conversely, a broad technology-specific exclusion may make attribution crucial even when it has limited predictive relevance to the underlying expected loss. This is a distinction between the evidence needed to forecast and the evidence needed to adjudicate.
Evaluation of coverage design
A policy can use explicit AI language to clarify a technological example within an existing definition, specify a class of AI-related insured interests, or introduce restrictions. Each option has consequences for comparability and customer expectations. Clarifying language can reduce some semantic disputes without expanding every covered loss. A restriction may preserve insurance capacity if it targets a material concentration and can be interpreted consistently. A restriction whose scope is difficult to observe may instead shift disputes into costly forensic inquiries and leave customers uncertain about protection.
The economic quality of a wording is connected to the verifiability of the fact upon which it conditions payment. When selection and pricing depend on the security measures of the insured, their implementation can often be observed through documentation and controlled testing. When coverage depends on an unobservable tool used by an unknown attacker, verification is more difficult. The appropriate scientific comparison includes the cost of that evidence and the resulting uncertainty of settlement, rather than considering coverage breadth in isolation.
Substantive protection should be assessed by scenario, considering multiple grants and their interaction. A high sublimit is of limited practical value when a narrowly defined insured event seldom encompasses the relevant economic loss. Conversely, a broad event definition with a small retention and a substantial sublimit may provide significant protection even though the aggregate policy maximum is lower than the organization's worst conceivable financial loss. Both insurer solvency and the insured's residual loss distribution matter.
A program for longitudinal empirical research
An improved dataset should link policy records and exposure periods to distinct incident records, then link each incident to notices, grants, payments and recoveries. Useful policy fields include insurer and form edition, endorsements, insured industry, size, jurisdiction, coverage dates, limits, waiting periods and retentions. Incident fields should include dates, mechanism, affected assets and providers, evidence of AI roles and confidence, safeguards in place, gross economic loss, and response actions. Claim records should identify amounts submitted, eligible or disputed, incurred, paid, recovered and outstanding.
The dataset must preserve provenance and privacy. Where sharing individual records is infeasible, a data steward could publish versioned aggregate tables with transparent inclusion criteria, minimum cell sizes and separate unknown-attribution counts. Duplicated incidents across insurers should be reconciled using suitable identifiers while preserving legitimate multiple contractual obligations. The presence of censorship or truncation due to reporting thresholds, retentions, limits and closure status should be explicit. A model fitted only to paid losses cannot be interpreted as a distribution of gross victim loss without adjustment.
A prospective empirical program would compare cohorts across industries and contract forms, estimate conventional claims frequency and severity, validate AI involvement classification with independent review, and study the relation between control implementation and covered outcomes. It should report the rate of missing technical information and uncertainty in legal classification, test alternative assumptions for common dependencies, and compare estimates on out-of-period observations. It would also assess whether new wording reduces dispute frequency or merely displaces disputes into other definitions. Establishing any claimed causal effect would require separately justified identification.
Boundaries of the present evidence
This investigation has examined primary policy texts in publicly distributed sample form, two U.S. judicial decisions, selected official intelligence and complaint reporting, an original Hong Kong incident account, selected market and guidance materials, and open actuarial literature. The examined materials establish meaningful distinctions and conditional mechanisms. They do not supply a representative international sample of issued policies or matched claim-level outcomes. Market reports aggregate heterogeneous exposures, and survey scenarios are elicited judgments rather than loss events.
Specific additional research remains necessary for a comprehensive jurisdictional account of binding law and for quantitative AI-specific underwriting estimates. Current EU legislative text retrieval was impeded, and the legal comparison therefore restricts its EU conclusions. Some full report texts were unavailable or only publicly indexed; the exact evidence scope is documented rather than imputed. The use of a dated NIST taxonomy is intentional, and no claim is made that it represents the latest revision. These limits also preclude a verified claim of scientific priority. The original contribution is a documentary comparison, causal interpretation, and explicit quantitative illustration grounded in established actuarial reasoning.
The practical research objective is to make insurance evidence more informative than a tally of purported AI attacks. A dataset that measures the conditional financial consequences of verifiable events under specified contracts can support meaningful economic inference. Data that merely classify losses by sensational descriptions risk conflating the threat environment with policyholder exposure and insurance payment.
Conclusion
Cyber insurance coverage of AI-enabled threats depends on the relationship between the technological activity, the resulting insured event, the defined financial loss, and the conditions of the insurance contract. Public threat assessments establish plausible and increasingly important mechanisms of attack assistance; documented incidents demonstrate actual economic harm; specimen policies reveal operative differences in coverage and authentication; judicial opinions illustrate the importance of contract language and factual causation. None supplies a universal AI-specific claim rate.
The documentary comparison establishes the relevance of technology-neutral grants and express synthetic-impersonation provisions, alongside separate limitations for verification, loss ownership, timing and aggregation. The economic analysis demonstrates that attacker capability may change both claim frequency and severity and that common dependencies can amplify portfolio tail exposure without a corresponding change in expected loss. The measurement illustrations establish why incomplete attribution and variable reporting can make observed trends misleading.
The integrated scientific conclusion is that insurance research should retain the underlying cyber peril as a stable classification, record AI's role and evidentiary confidence as additional dimensions, and determine the economic and legal response from the exact insured interests and operative clauses. This approach supports valid comparisons through technological change, more transparent coverage assessment, and more defensible portfolio analysis. It also identifies the specific data and validation requirements for future empirical estimates.
Insurance is an institutional mechanism for financing defined contingent loss. Its capacity to accommodate evolving AI-enabled cyber threats depends upon meaningful contractual specification, evidence of actual events and economic harm, adaptive risk assessment, and explicit treatment of dependence. The protection of policyholders and the sustainability of insurance capacity are linked through the quality of those practices.
References
American Tooling Center, Inc. v. Travelers Casualty & Surety Co. of America, 895 F.3d 455 (6th Cir. 2018), No. 17-2014, July 13. Opinion reproduced by Justia. Source
At-Bay. (2023). Cyber insurance policy. Form AB-CYB-001.2, edition 08/2023. Source
Awiszus, K., Knispel, T., Penner, I., Svindland, G., Voß, A., & Weber, S. (2023). Modeling and pricing cyber insurance: Idiosyncratic, systematic, and systemic risks. European Actuarial Journal, 13, 1–53. doi:10.1007/s13385-023-00341-9. Source
AXIS. (2020). AXIS cyber insurance policy (specimen). Form AXIS 1012561 0120. Source
Coalition. (2026). 2026 cyber claims report: Public key findings and report landing page. Source
Coalition. (n.d.). Cyber sample policy (CIC). Sample declarations and policy form distributed by Victor Insurance. Source
European Securities and Markets Authority (ESMA). (n.d.). Digital Operational Resilience Act (DORA). Official implementation overview of Regulation (EU) 2022/2554. Source
Federal Bureau of Investigation (FBI), Internet Crime Complaint Center (IC3). (2026). 2025 IC3 annual report. FBI. Source
Government of the Hong Kong Special Administrative Region. (2024, June 26). LCQ9: Combating frauds involving deepfake. Legislative Council written reply. Source
Lloyd's Market Association (LMA). (n.d.). Cyber war clauses. Clause classification and compliance guidance following Y5381 and Y5433. Source
Lloyd's Market Association (LMA). (n.d.). Understanding AI exposures: AI loss scenarios survey results. Survey issued April 2025; wording review updated through November 2025. Source
Lloyd's. (2022, August 16). State backed cyber-attack exclusions. Market Bulletin Y5381. Source
Medidata Solutions Inc. v. Federal Insurance Co., 729 F. App'x 117 (2d Cir. 2018), No. 17-2492-cv, July 6. Summary order reproduced by FindLaw. Source
National Association of Insurance Commissioners (NAIC). (2026). Report on the cybersecurity insurance market. Calendar-year 2025 analysis. Source
National Cyber Security Centre (NCSC). (2025). Impact of AI on cyber threat from now to 2027. UK National Cyber Security Centre. Source
Pascoe, C., Quinn, S., & Scarfone, K. (2024). The NIST cybersecurity framework (CSF) 2.0. NIST CSWP 29. doi:10.6028/NIST.CSWP.29. Source
U.S. Department of the Treasury, Office of Foreign Assets Control (OFAC). (2021, September 21). Updated advisory on potential sanctions risks for facilitating ransomware payments. Source
Vassilev, A., Oprea, A., Fordyce, A., Anderson, H., Davies, X., & Hamin, M. (2025). Adversarial machine learning: A taxonomy and terminology of attacks and mitigations. NIST AI 100-2e2025. doi:10.6028/NIST.AI.100-2e2025. Source